【作者】 朱娟；

【导师】 王延杰；

【作者基本信息】 中国科学院研究生院（长春光学精密机械与物理研究所） ， 机械电子工程， 2010， 博士

【摘要】 准确,鲁棒的跟踪性能是跟踪学者们所追求的目标,为此,本论文选择了对目标模型和噪声没有限制的蒙特卡洛滤波算法。通过查阅国内外文献及软件仿真,对蒙特卡洛滤波算法的原理进行了深入的分析和研究,针对传统的蒙特卡洛滤波算法存在的主要问题——计算复杂度大和样本退化问题,分别提出了改进措施。好的模型是跟踪成败的关键,针对传统模型不能去除背景干扰的问题,论文提出了新的目标表观建模方案——基于轮廓的表观模型提取。首先通过边缘检测,直线段检测等一系列步骤提取出给定区域内目标的轮廓,用轮廓曲线内的信息表示目标,接下来用颜色信息和形状信息相结合的方式描述目标特征,使得特征描述符可以抵抗光照变化及旋转缩放等复杂情况。实验结果表明,在PC机上用VC++6.0实现时,对于300*300像素的区域以内的目标,这种建模方法提取模型时间<1ms,几乎可以忽略,匹配过程中,用新模板的匹配时间平均可以达到用原模板匹配时间的1/2。针对蒙特卡洛算法的样本退化问题,论文提出了基于人眼机制的样本集建立方案和半采样半重采样的样本集传播策略。初始帧需要建立样本集合,按照与中心距离越近样本越密集,反之越稀疏的原则建立样本集合;样本传播过程中,将样本集合分成优劣两类,对优样本集合通过重采样传播,对劣样本集合用样本建立方法重新采样代替。实验结果表明,这种样本集建立传播策略解决了样本的退化问题。针对Mean Shift算法不能很好解决非线性目标跟踪以及蒙特卡洛算法实时性差的问题,提出了一种自适应选择Mean Shift算法与蒙特卡洛算法的目标跟踪策略。引入一个跟踪方式选择标志,通过计算当前跟踪窗内的目标与模板的相似度来决定该标志的值,当目标与模板的相似度大于某个阈值时,选择实时性好的基于梯度最速下降的Mean Shift跟踪策略,以实现跟踪的实时性;否则选择基于随机采样、对目标模型没有限制的蒙特卡洛跟踪策略,使得位置预测结果更加准确。实验结果表明:与蒙特卡洛算法相比,该算法在跟踪性能不受影响的前提下,有效节省了系统时间,当目标简单运动时,对于100*56像素的目标,平均计算时间由原来的82ms降低为小于1ms;与Mean Shift算法相比,该算法在牺牲一些系统时间的基础上能够更加鲁棒地解决非线性目标跟踪问题。更多还原

【Abstract】 To track target accurately and robustly, this paper chooses Monte Carlo filtering algorithm which is free from target motion model and noise. Have been referred to a great deal of overseas or domestic literatures, and many simulator experiments, deeply analyze and research on the fundamental theories of Monte Carlo Filtering. Traditional Monte Carlo Filtering has two fatal problems——the large computational load and samples degeneracy. To solve these two problems, new methods are presented.Good model has a key position on tracking problem. Traditional model could not resist the interference of background. So, a new model building method which is based on the outline is proposed. Firstly, get the target outline which is in a given region through a set of steps such as edge detection, line detection and so on. Describe the target using the information which is inside the outline. Secondly, describe target feature combining color and shape information. So the feature description can resist the change of backlight and target itself. Experimental results show the time of building model less than 300*300 pixels is less than 1 millisecond using VC++, which could be almost ignored. The matching time using new model is greatly decreased than that of using original model.Traditional Monte Carlo filtering algorithm has the problem of sample degeneracy. To solve this problem, this paper proposed samples building method based on visual principle and samples propagation method using half sampling and half re-sampling. At original frame, build the samples according to the distance from the center. Get denser samples where near the center and get sparser samples where far away from the center. During the samples propagation, classify the samples into good samples and bad samples. Re-sample the good samples for propagating and get new samples to replace the bad samples. Experimental results show the sample degeneracy problem is solved by this method.Mean Shift deals badly with non-linear problems and Monte Carlo method increases the computational load greatly. In order to solve these problems, a new target tracking method that chooses Mean Shift and Monte Carlo method to track object adaptively is proposed. Introduce a sign that denotes the tracking method to this article for choosing Mean Shift and Monte Carlo. The sign is confirmed by the match degree of current target and the model. When the match degree is above a given threshold, real- time Mean Shift method which is based on gradient method is chosen to track the target fast. Otherwise, Monte Carlo method which is based on random sampling and free from target model is chosen to track the target exactly. The basic theories and the whole frame of the tracking system are given. Experimental results prove that compared with Monte Carlo, proposed method has the same tracking performance and costs less time. When the target moves linearly, the average time is 82 milliseconds for Monte Carlo and is less than 1 millisecond for proposed method when the target is 100*56 pixels. Compared with Mean Shift, proposed method is more robust to non-linear problems.更多还原